What Is MCP? The Model Context Protocol Explained for Builders

MCP, or Model Context Protocol, is an open-source standard for connecting AI applications to external systems. It gives assistants like Claude and ChatGPT one shared way to read context, such as files or metrics, and to call tools, such as running a query. Build one MCP server, and any compatible AI app can use it.

The official docs compare MCP to a USB-C port for AI applications. Before USB-C, every device came with its own cable. Before MCP, every AI app needed its own connector for every system it touched.

MCP exists because every pairing of an AI app and a data source used to need its own custom integration. Developers call this the N×M problem. Five AI apps connecting to 10 tools means 50 separate integrations, each built and maintained on its own.

Anthropic named the problem directly when it published its MCP announcement on Nov. 25, 2024:

"Every new data source requires its own custom implementation, making truly connected systems difficult to scale."

MCP turns N×M into N+M. Each AI app implements the client side once, and each tool implements the server side once. In the example above, that drops 50 integrations to 15.

MCP works by placing a client inside your AI app that talks to one or more MCP servers. Each server describes what its system can do, the model picks the right tool for your question, and the server returns structured results that the model turns into an answer.

Hosts, clients, and servers

Four parts make up every MCP connection:

• Host: The AI app you're using, such as Claude Code, Cursor, or ChatGPT.
• Client: The connector inside the host. The host runs one client per server.
• Server: A program that exposes a system, such as an analytics platform or a code repository, through MCP.
• Transport: How messages travel between client and server, either locally on your machine or over HTTP for remote servers.

Here's how one request flows for a hypothetical B2B SaaS team investigating onboarding:

1. You ask Cursor, "Why did activation drop last week?"
2. Each client has already asked its server which tools it offers, so the model knows a funnel query tool exists.
3. The model picks that tool, and the client sends the call to the analytics server.
4. The server runs the query with your permissions and returns structured results.
5. The model reads those results and writes an answer grounded in them.

Tools, resources, and prompts

MCP servers offer three kinds of building blocks:

• Tools: Actions the model can call, like running a funnel query or opening a ticket.
• Resources: Data the app can read for context, like a file or a database schema.
• Prompts: Reusable templates a server provides, like "summarize this experiment."

What is an MCP server?

An MCP server is the piece that connects one system to any MCP-compatible AI app. It translates the protocol's standard requests into that system's own queries or API calls, then sends results back in a format the model can use.

MCP server examples

Common MCP servers cover the systems builders already use every day. You'll find servers for code repositories, issue trackers like Jira, documentation tools like Confluence, and analytics platforms. Combining them is where MCP pays off: one session can read a metric drop, find the related release in your repo, and open a ticket.

Analytics platforms ship servers too: Amplitude MCP gives MCP-compatible clients that support remote HTTP servers and OAuth access to your product data. You can also connect it through Figma, GitHub, Kiro, Notion, Rovo, and other AI applications.

MCP, APIs, and RAG solve different problems, and they often work together. An application programming interface (API) exposes a system to code. Retrieval-augmented generation (RAG) feeds a model relevant documents from an index. MCP gives AI apps a standard way to discover and use live tools and data at runtime, often by wrapping existing APIs.

A simple rule of thumb: use RAG when the answer lives in documents, and use MCP when the answer lives in a system. "What's our refund policy?" is a RAG question. "How many users hit the paywall last week?" is an MCP question.

Most major AI assistants and coding tools support MCP. The official docs list Claude, ChatGPT, VS Code, and Cursor as supporting applications. Claude Code connects to servers from the terminal, and Amplitude MCP works with Codex. For GitHub Copilot, confirm current support in GitHub's documentation.

• Claude and Claude Code: Anthropic created the protocol, so these are natural first clients to try.
• ChatGPT: Yes, ChatGPT uses MCP, though setup steps differ from other clients.
• Cursor: Listed in the official docs as a supporting application.
• VS Code and GitHub Copilot: VS Code appears in the official docs' list.
• Codex: Amplitude MCP works with Codex, and you can connect Amplitude MCP to Codex CLI as a starting point.

MCP analytics is the practice of connecting AI assistants to an analytics platform through an MCP server. Questions about activation, conversion, or retention then get answered from your real event data, using the same governed definitions your dashboards use, instead of the model's best guess from training.

That matters because language models fill gaps with confidence. A widely cited survey of LLM hallucination by Lei Huang and colleagues, published in ACM Transactions on Information Systems in 2025, warns that "LLMs are prone to hallucination, generating plausible yet nonfactual content." The authors also note that a model's knowledge isn't updated after training. Ask an unconnected assistant why activation dropped, and you'll get a reasonable-sounding story with no data behind it.

Grounding an assistant in governed data reduces the risk of a confident wrong answer, but it doesn't remove it. The same survey points out that retrieval-augmented models can still hallucinate.

Amplitude MCP exposes Amplitude analytics, taxonomy, and content, including charts, dashboards, experiments, cohorts, events, and properties, as tools that compatible AI clients can call over OAuth. Through it, you can query analytics, run ad-hoc segmentation, funnels, and retention, and analyze experiment results. For example, you could ask:

• Activation: "Why did activation drop last week?"
• Conversion: "Show me how last week's release impacted signup funnel conversion."

Amplitude also offers AI Agents, which run analyses for you. Either way, the assistant can only see what you track, so if an event isn't instrumented, no prompt will surface it.

Zip has seen this firsthand. Moss Pauly, technical product manager for data and AI at Zip, explains on the Amplitude MCP page:

"We've been able to embed our domain knowledge directly into our LLM workflows. Now insights that used to require specialized analytics expertise are accessible to everyone."

Using MCP securely comes down to three things: how the server authenticates you, what it can do on your behalf, and which client runs it. The protocol doesn't make you safe by default, so you need scoped permissions, approval for write actions, and servers from sources you trust.

The Journal of Cybersecurity and Privacy study tested seven clients against tool poisoning, where a malicious server hides instructions in its tool descriptions. Success rates ranged from 0% in Claude Desktop to 100% in Cursor, and five of the seven clients skipped static validation of tool metadata. That research targets malicious servers, which is a different risk from a first-party server run by a vendor you already trust.

Practices that hold up:

• Prefer OAuth over shared API keys. The MCP specification builds its authorization rules on OAuth and tells clients to follow the principle of least privilege.
• Inherit permissions instead of granting new ones. A server acting for you should see only what you can already see.
• Separate read tools from write tools, and approve writes before they run.
• Know where a server runs. Local servers run with your machine's access, while remote servers run on the provider's infrastructure.
• Install servers only from sources you trust, and review tool descriptions before approving them.

Amplitude's server applies the permission rule directly: it uses your existing Amplitude permissions, so it only sees projects you can already access. The Amplitude MCP docs explain setup and OAuth authentication.

Connect Amplitude MCP and ask your first product question from the AI tool you already use. In Claude Code, setup takes about two minutes. You add the server, authenticate with OAuth, and confirm the connection by asking which Amplitude projects you can access.

Start with the guide to use Amplitude MCP with Claude Code, or follow the docs to connect Amplitude MCP to Cursor or Codex CLI. For background on why Amplitude built it, read the Amplitude MCP launch post.

Every Amplitude plan includes Amplitude MCP, the Free plan too, so you can start with the account you have.